cloudg reference

cloudg (cloud graphing) maps AWS, Azure and GCP infrastructure into a graph, runs security scanners over that inventory, and merges every finding into one deduplicated, compliance-mapped result set. It can also skip the scanning entirely and aggregate outputs you already have.

This is the complete reference: every CLI command, the exact input and output formats, the data models, the full Python API with integration examples, configuration, plugins and authentication. Release history lives in the changelog.

The inventory mapper (cloudg map, cloudg deps, cloudg.inventory) has three deeper companion documents:

• Inventory reference: the structure of every return value and exported file, field by field, with annotated examples (CloudAsset, NetworkEdge, relation objects, InventoryResult, summary, coverage records, dependency trees, blast radius, security coverage, inventory-map.json, inventory-graph.json, GraphML, inventory-organization.json, asset-map.json, compliance-map.json, cloudg deps --json).
• Inventory catalog: all 156 asset types with the native resource types mapped to each, the metadata keys and relations each one carries, and the full relationship matrix.
• Inventory internals: how a mapping run works, every module, the 133 AWS collector tasks, the linker's resolution rules, the Azure and GCP extractor frameworks, dependency semantics, catalogs, and recipes for adding collectors, extractors and asset types.

Since 0.6.0 cloudg also runs as an MCP server for AI agents, and paces every cloud API call it makes. Both have their own documents on GitHub:

• MCP guide: installing and serving the MCP layer, client configuration, transports and protocol versions, mounting it into an existing MCP server, in-process use, middleware, the live tools.
• MCP tool reference: every tool, resource, resource template and prompt, with arguments and result shapes.
• MCP privacy: policies, profiles, roles, transforms, detectors and the pseudonym vault.
• MCP internals: how the layer is built, for contributors.
• Rate limits and throttling: the shared rate limiter, circuit breakers and retries, per-provider wiring, every ratelimit key, and the guard on live collections.

Contents:

• Installation
• The pipeline
• CLI reference
• Input requirements
• Output contract
• Data models
• Configuration
• Python API
• Integration recipes
• MCP server
• Rate limits and throttling
• Plugin system
• Authentication
• Docker

Installation

From PyPI:

pip install cloudg            # core
pip install cloudg[aws]       # + boto3, aioboto3
pip install cloudg[azure]     # + Azure management SDKs
pip install cloudg[gcp]       # + google-cloud-asset, google-auth
pip install cloudg[mcp]       # + the official MCP SDK (mcp>=1.30,<3), see MCP server
pip install cloudg[all]       # every provider SDK and the MCP SDK

The cloud SDKs are optional extras, pulled in only when you collect from that provider. The core package alone ships the graph engine, the ontology builder, the findings normaliser, the report renderers and cloudg ingest. That is enough to aggregate existing scan results on a machine with no cloud access at all.

The external scanners are separate executables, not Python dependencies: Prowler (pip install prowler), ScoutSuite (pip install scoutsuite), Checkov (pip install checkov) and Trivy (binary install). Install whichever subset you want. cloudg checks PATH at run time and skips anything missing with a warning; the rest of the pipeline is unaffected. install.sh (Linux/macOS) and install.bat (Windows) set up all of them plus the cloud CLIs, and the Docker image bundles the four scanner binaries.

Requires Python 3.11 or newer.

The pipeline

collect -> graph -> scan -> normalise -> report
                \-> ontology / RAG / terraform

Collection runs all configured providers concurrently with asyncio, iterating accounts and regions per provider. Regions auto-discover when you pass --regions all. Assets and network edges go into a directed NetworkX graph: BFS from the internet node finds exposed resources, and blast-radius scoring estimates what an attacker could reach from each node.

The enabled scanners then run in parallel over the same inventory, each in its own thread with its own timeout. Their findings, plus the graph's own reachability findings and the built-in IAM linter's results, all flow into the normaliser, which deduplicates within and across scanners, rescores against CVSS, and maps everything to compliance controls. The renderers turn the final ScanResult into the report files listed under Output contract.

Every phase degrades gracefully. A missing scanner binary, an unreachable provider or a failed exporter logs a warning; nothing else stops.

Alongside the pipeline sits an independent function: inventory mapping (cloudg map / CloudGEngine.map_inventory()). It runs no scanners at all. It deep-collects everything deployed or default in the account, sweeps every service for whatever the dedicated collectors miss, and links the lot into one map. Scanner findings can be merged into that map later.

CLI reference

Global options go before the subcommand: cloudg -v -c my-config.yaml run ...

Option	Meaning
-v, --verbose	debug logging
-c, --config PATH	path to config.yaml
--log-file PATH	also write logs to a file

cloudg run

The full pipeline: collect, graph, scan, normalise, report.

cloudg run -p aws --regions us-east-1
cloudg run -p all --regions all
cloudg run -p aws --scanners prowler,trivy
cloudg run -p aws --regions us-east-1 --terraform

Flag	Meaning
-p, --provider	aws, azure, gcp or all; repeatable
--regions	all for auto-discovery, or a comma-separated list
--scanners	subset of prowler,scoutsuite,checkov,trivy,iam; defaults to scanners.enabled in config
--iac-dir	directory for the IaC scanners (Checkov, Trivy fs)
--images	comma-separated container images for Trivy
--ontology / --no-ontology	RDF export, on by default
--rag-export / --no-rag-export	RAG chunks, on by default
--terraform / --no-terraform	Terraform recreation, off by default
-o, --output	output directory, ./reports by default

Credential flags are listed under Authentication.

A detail worth knowing: when no IaC directory is configured but --terraform is on, the IaC scanners target the generated Terraform recreation of the live infrastructure. Checkov then audits your actual cloud rather than whatever directory cloudg happens to run from. There is deliberately no fallback to .: a zero-finding scan of an unrelated directory reads like a clean bill of health, and cloudg refuses to produce one.

cloudg collect

Asset collection alone, one provider at a time. Writes assets and edges without running any scanner.

cloudg collect -p aws --profile prod --region eu-west-1 -o ./reports
cloudg collect -p azure --subscription-id <id>
cloudg collect -p gcp --project-id <id>

cloudg map

Scanner-independent inventory mapping: everything deployed or default, linked into one map of typed interdependencies. No scanner runs and none needs to be installed.

cloudg map -p aws --regions all
cloudg map -p all --regions all
cloudg map -p aws --org --regions all                       # every account in the Organization / Control Tower
cloudg map -p aws --org --ou Workloads --exclude-account 111122223333
cloudg map -p aws --services containers,serverless,security  # narrow the run
cloudg map -p aws --findings ./reports/raw-findings.json     # merge scanner output

Flag	Meaning
-p, --provider	aws, azure, gcp or all; repeatable
--regions	all for auto-discovery, or a comma-separated list
--profile	AWS CLI profile
--accounts, --role-name	AWS: explicit account list and the role to assume in each
--org / --no-org	AWS: discover and map every account of the Organization (run from the management account or a delegated administrator)
--org-role	AWS: role assumed in member accounts; defaults to AWSControlTowerExecution
--ou	AWS: only accounts under this OU (ID, ARN or name, nested OUs included); repeatable
--exclude-account	AWS: skip an account; repeatable
--ct-home-region	AWS: Control Tower home region, auto-detected when omitted
--services, --exclude-services	service families or collector names, comma-separated (see below)
--kubernetes / --no-kubernetes	map workloads inside EKS clusters through the Kubernetes API
--cloud-control / --no-cloud-control	AWS Cloud Control breadth sweep, on by default
--subscription-id, --project-id	Azure / GCP identity
--findings	existing cloudg findings JSON (raw-findings.json or a report); repeatable
--sweep / --no-sweep	catch-all sweep, on by default
-o, --output	output directory

Coverage comes in four layers.

1. Dedicated deep collectors. On AWS, 133 collectors grouped into service families you can select with --services / --exclude-services (family or collector names):

   Family	Collected
   network	VPCs, subnets, security groups, route tables, internet / NAT / egress-only gateways, ENIs, Elastic IPs, NACLs, VPC peering, transit gateways, attachments, route tables and peering, VPC endpoints and endpoint services (PrivateLink, allowed principals, consumers), managed prefix lists, ALB/NLB listeners, rules, certificates and target groups, classic ELB, CloudFront (origins, aliases, certificates, Lambda@Edge, functions, OAC, logging), Global Accelerator, VPC Lattice
   hybrid	site-to-site VPN, customer and virtual private gateways, Direct Connect connections, VIFs and gateways, Cloud WAN / Network Manager
   compute	EC2, Auto Scaling groups, launch templates, AMIs (with sharing), Batch, App Runner, Elastic Beanstalk
   containers	ECR (+ public), ECS clusters, services, task definitions, capacity providers, container instances, Cloud Map, EKS clusters, nodegroups, Fargate profiles, addons, access entries, pod identity
   kubernetes	inside EKS clusters: namespaces, Deployments, StatefulSets, DaemonSets, CronJobs, Services, Ingresses, ServiceAccounts
   serverless	Lambda (triggers, URLs, images, layers, DLQs, aliases), API Gateway REST / HTTP APIs, authorizers, VPC links and custom domains, Step Functions
   integration	SQS, SNS, EventBridge buses, rules, archives, API destinations, Scheduler, Pipes, Kinesis, Firehose, MSK, Amazon MQ
   storage	S3 (notifications, replication, logging, account and bucket Block Public Access), access points and multi-region access points, EBS volumes and snapshots, EFS, FSx, Transfer Family, DataSync
   data	RDS, Aurora, proxies, global clusters, manual snapshot sharing, DynamoDB (KMS, streams, replicas, PITR, policies), ElastiCache, MemoryDB, DAX, OpenSearch (+ Serverless), Redshift (+ Serverless)
   analytics, ml	Glue, Lake Formation permissions, EMR (+ Serverless), Athena; SageMaker, Bedrock agents, knowledge bases, guardrails and invocation logging
   identity	the IAM graph (users, groups, roles, trust, grants, instance profiles, OIDC and SAML providers, access key metadata), IAM Identity Center (instances, permission sets, assignments, users, groups), Roles Anywhere, Cognito user and identity pools, KMS (aliases, rotation, key policy and grant principals), Secrets Manager, ACM
   governance	RAM resource shares, Service Catalog portfolios and provisioned products (Account Factory), CloudFormation StackSets
   security	GuardDuty, Security Hub, Inspector (per-resource coverage), Macie, AWS Config, IAM Access Analyzer, Detective, WAF, Network Firewall, Shield Advanced
   logging, operations, backup	CloudTrail, flow logs, log groups, subscription filters and destinations, cross-account observability (OAM); CloudWatch alarms, SSM parameters (metadata only), managed instances, documents, associations, maintenance windows; AWS Backup plans, selections and vaults
   dns, iac, cicd	Route 53 zones and records, Resolver endpoints, rules and DNS Firewall; CloudFormation stacks and the resources they manage; CodePipeline, CodeBuild, CodeDeploy, CodeConnections

   Account-wide services are collected once per account, in the primary region, so --regions all does not repeat them. Security services that are not enabled appear as placeholder nodes with enabled: false, so detection gaps show on the map. Values that could be secret (parameter values, environment variables, passwords, VPN pre-shared keys, Direct Connect auth keys) are never collected.

   Azure is collected through Azure Resource Graph (azure-mgmt-resourcegraph, in the azure extra) with every resource's full properties, so PaaS services get real relationships: managed identities and role assignments (principal → scope, with role names), private endpoints, AKS node pools, kubelet identity and ACR pulls, App Service plans, VNet integration and container images, Container Apps, subnets, peering, NICs, disks and disk encryption sets, firewalls and route-table next hops, Application Gateway and Front Door backends and WAF policies, Key Vault access policies, storage and SQL network rules, Defender for Cloud plans per subscription. Every Enabled subscription is collected when none are configured (azure.all_subscriptions), and the management group → subscription → resource group hierarchy, Azure Landing Zone archetypes and Azure Policy assignments and exemptions are mapped (azure.map_management_groups). Without the Resource Graph SDK, cloudg falls back to the per-service SDK collectors plus the ARM sweep. Permissions: Reader at the tenant root management group (Resource Graph scopes to what the credential can read).

   GCP is collected through Cloud Asset Inventory list_assets with the full resource JSON, once at organizations/<organization_id> when that is set (gcp.collection_scope), otherwise per project. Relations cover instances (service accounts, subnets, disks, external IPs), firewall semantics (target tags and service accounts, source ranges, priorities), the load-balancing chain down to NEGs and Cloud Armor, managed instance groups and templates, GKE clusters and node pools with Workload Identity, Cloud Run and Functions v2 (service accounts, images in Artifact Registry, secrets, VPC connectors, triggers), Pub/Sub (subscriptions, push endpoints, dead-letter and export targets), Eventarc, logging sinks, CMEK keys, Shared VPC, Cloud SQL / Redis / Filestore networks and DNS zones. IAM bindings resolve to service accounts or to principal nodes (users, groups, workload identity pools); only allUsers / allAuthenticatedUsers mark a resource internet-exposed. With an organization ID, the organization → folder → project tree, organization policies and VPC Service Controls perimeters are mapped too (gcp.map_hierarchy, org_policies, vpc_service_controls). Permissions: roles/cloudasset.viewer on the organization (or each project) and the Cloud Asset API enabled on the quota project.

2. Breadth sweeps, so services without a dedicated collector still appear on the map.

   • AWS Cloud Control API: every CloudFormation resource type with a list handler (800+ types) is listed, tagged or not. Types already mapped by a deep collector are skipped, account-wide types run once per account, types that need a parent identifier or fail are skipped and counted in coverage (cloud_control_skipped_types). The listable-type catalogue is discovered once and cached for a week ($CLOUDG_CACHE_DIR, default ~/.cache/cloudg). Secret-bearing properties (passwords, tokens, SSM parameter values, environment variables, user data) are never copied. Turn it off with --no-cloud-control, or narrow it with inventory.cloud_control_types / cloud_control_exclude (type names or prefixes such as AWS::Glue::).
   • AWS Resource Groups Tagging API: tagged resources only (AWS never returns resources that were never tagged), kept for tag enrichment.
   • Azure Resource Manager's full resources.list() and GCP Cloud Asset Inventory enumerate their scopes.

   Precedence when the same resource is found twice: dedicated collector, then Cloud Control, then the tagging API.

3. The relationship linker. Collectors declare what each asset talks to; the linker resolves those identifiers across services, regions and accounts into typed edges, then applies provider rules and a generic reference pass. See Relationships below.

4. Account hierarchy. Every account gets a CLOUD_ACCOUNT node that contains its top-level resources. With --org, accounts hang under the OU tree, next to SCPs, the Control Tower landing zone and its enabled controls.

Outputs: inventory-map.json (assets, edges, unresolved references, and a summary covering services, types, regions, accounts, relationship counts, cross-account edges, external accounts, security service gaps, internet exposure and unlinked assets), inventory-map.graphml, inventory-graph.json (D3, with account and relationship attributes), inventory-dependencies.json (most shared dependencies, largest blast radius, cross-account edges, security coverage including workloads no vulnerability scanner covers and internet-facing endpoints without a WAF), and with --org inventory-organization.json. With --findings, additionally asset-map.json (per-asset finding counts and severity breakdowns, riskiest first) and compliance-map.json (framework → findings and affected assets). The map does not depend on the scanners; findings from any earlier run merge in whenever they exist.

Relationships

Every edge has a coarse edge_type and, when known, a fine-grained relationship (an ontology relation name) plus properties. Direction always reads "source verb target".

Edge type	Example
INVOKES	S3 bucket → Lambda (notification), SQS → Lambda (TRIGGERED_BY, event source mapping), SNS → SQS (STREAMS_TO), EventBridge rule → target, API Gateway → Lambda, Step Functions → anything its definition calls
USES_IMAGE	task definition / Kubernetes workload / Lambda → ECR repository
ASSUMES_ROLE	Lambda, task definition, nodegroup, Kubernetes service account (IRSA or Pod Identity), instance profile → IAM role
IAM_TRUST	principal / external account / OIDC provider → role it may assume (CROSS_ACCOUNT_TRUST across accounts)
GRANTS_ACCESS	principal → resource its policy names; principal → EKS cluster (access entry, with access policies); bucket / queue / repository policy grants
IAM_POLICY_ATTACHMENT	user, group or role → managed policy
PROTECTS	WAF → ALB / API Gateway / CloudFront, Network Firewall → VPC, Shield → resource
MONITORS	Inspector → each scanned EC2 / ECR / Lambda, GuardDuty / Config / CloudTrail / Macie → account, Security Hub → the services it ingests from, flow log → VPC/subnet/ENI
MANAGES	CloudFormation stack → resource, ASG → instance, nodegroup → ASG, landing zone → OUs and shared accounts
GOVERNS	SCP / RCP / tag policy and Control Tower control or baseline → OU or account
LOGS_TO	trail, flow log, load balancer, API stage, task definition, function → log destination
LOAD_BALANCER_TARGET	load balancer → target group → instance / IP / Lambda, Kubernetes Service → workload
ROUTE	route table → gateway, DNS record → load balancer / distribution, transit gateway → attached VPC, Ingress → Service
CONTAINS, ATTACHED_TO, PEERING, REFERENCES	containment, security-group attachment, VPC peering, everything else (KMS keys, secrets, DLQs, layers, ...)

Identifier resolution is scoped: an ambiguous name such as default resolves within the referencing asset's account and region, and is left unresolved rather than guessed when it could belong to several accounts. References into accounts that were not mapped become CLOUD_ACCOUNT nodes with external: true, so a role trusted by a vendor account or a bucket replicating to another account stays visible. References that resolve to nothing (a deleted queue still configured as a DLQ) are listed under unresolved_references.

cloudg deps

Interdependency queries over a saved map; no cloud access needed.

cloudg deps arn:aws:iam::123456789012:role/app-role        # both directions
cloudg deps jobs-queue --direction down --depth 5           # blast radius
cloudg deps --map ./reports                                 # overview
cloudg deps my-function --direction up --json               # machine-readable

Each edge is read as "dependent needs dependency": a function needs its role, image, key and the queue that triggers it; an ALB needs the WAF protecting it; a resource needs the stack managing it. --direction up shows what an asset needs, down what needs it (what breaks or changes with it), both the two trees. Without an asset, deps prints the most shared dependencies (one KMS key behind forty resources), the largest blast radius, and every cross-account edge. Assets are matched by ARN or resource ID, internal ID, a unique name, or a unique ARN tail (the part after the last /).

cloudg scan

Scanners alone, no collection. Writes raw-findings.json.

cloudg scan -p aws --scanners prowler,checkov --iac-dir ./terraform

cloudg ingest

Aggregation without execution. Feed cloudg the native output files of scans that already ran (in CI, on another host, on a schedule) and it runs the normalisation half of the pipeline: cross-scanner dedupe via the check-equivalence rulesets, compliance mapping and report generation. No scanner executes and no cloud credentials are needed.

cloudg ingest \
  --prowler ./prowler-output/ \
  --scoutsuite ./scoutsuite-report/ \
  --checkov ./results_json.json \
  --trivy ./trivy-image.json --trivy ./trivy-fs.json \
  -o ./reports

Flag	Accepts
--prowler	ASFF JSON or JSONL file, or Prowler's -o output directory (searched recursively for *.json)
--scoutsuite	the scoutsuite_results_*.js file, or the --report-dir directory
--checkov	a checkov --output json file (or results_json.json), or a directory containing it
--trivy	a trivy image or trivy fs JSON file, or a directory of them
--format	html, json or all (default)
-o, --output	output directory

Every input flag is repeatable and any combination or subset of tools works: one tool alone, or all four together. See Input requirements for the exact formats.

cloudg report

Re-render reports from a previous run's findings.json. No cloud, no scanners.

cloudg report -i ./reports/findings.json --format html

--format takes html, json, svg or all.

Input requirements

This section pins down exactly what each ingest path expects, and the command that produces it.

Prowler

Produce:

prowler aws -M json-asff -o ./prowler-output

cloudg accepts the output directory itself or any single file from it. Files may be a JSON array or JSONL (one ASFF object per line); both parse. The fields cloudg reads from each ASFF finding:

{
  "Id": "prowler-aws-s3_bucket_default_encryption-123456789012-eu-west-1-...",
  "Title": "S3 bucket default encryption",
  "Description": "...",
  "Severity": {"Label": "HIGH"},
  "Resources": [{"Id": "arn:aws:s3:::my-bucket"}],
  "Compliance": {"Status": "FAILED", "RelatedRequirements": ["CIS 2.1.1"]},
  "Remediation": {"Recommendation": {"Text": "..."}}
}

Findings with Compliance.Status of PASSED are skipped. The check name embedded in Id (the third dash-separated token) becomes the dedupe key.

ScoutSuite

Produce:

scout aws --report-dir ./scoutsuite-report --no-browser

cloudg accepts the report directory (searched recursively for scoutsuite_results*.js) or the file itself. The file is JavaScript, a single JSON object assigned to a variable; cloudg strips the assignment and parses the JSON. Findings come from services.<service>.findings.<key>, where each entry with flagged_items > 0 yields one finding per item in items. The level field maps to severity: danger becomes CRITICAL, warning HIGH, caution MEDIUM.

Checkov

Produce:

checkov -d ./iac --output json > results_json.json

cloudg accepts the JSON file or a directory containing results_json.json (falling back to any *.json in the directory). Both shapes Checkov emits parse: a single {check_type, results} object, or a list of them when several frameworks ran. Findings come from results.failed_checks[], reading check_id, check_name, severity, resource, file_path, file_line_range and guideline.

Trivy

Produce either or both:

trivy image --format json myrepo/app:latest > trivy-image.json
trivy fs --format json --scanners vuln,misconfig,secret ./iac > trivy-fs.json

cloudg accepts a single JSON file or a directory of them. Each report's top-level ArtifactType decides the parser: container_image goes through the image parser (vulnerabilities and secrets), everything else (filesystem, repository) goes through the filesystem parser, which also handles misconfigurations. ArtifactName becomes the resource identifier. From each vulnerability cloudg reads VulnerabilityID, PkgName, InstalledVersion, FixedVersion, Severity, Description and the CVSS v3 score when present.

Output contract

All commands write into the output directory (./reports by default).

File	Contents
report.html	interactive report: D3 topology, findings table, compliance matrix. Self-contained, works offline.
findings.json	the machine-readable result, structure below
raw-findings.json	pre-normalisation findings (from scan and ingest)
topology.svg, topology.graphml, topology-cytoscape.json	the graph in three formats
ontology.ttl, ontology.jsonld	RDF ontology, 64 inferred relation types, SPARQL-queryable
rag_chunks.jsonl, rag_metadata_index.json	retrieval-ready chunks, one JSON object per line
terraform/*.tf.json, terraform/import.sh	Terraform recreation of live infrastructure, 25+ asset types
inventory-map.json, inventory-map.graphml, inventory-graph.json	scanner-independent inventory map: assets, interconnections, summary (cloudg map)
inventory-dependencies.json	shared dependencies, blast radius, cross-account edges, security service coverage (cloudg map)
inventory-organization.json	AWS Organization / Control Tower topology (cloudg map --org)
asset-map.json, compliance-map.json	inventory overlaid with scanner findings (cloudg map --findings)

The inventory files are specified field by field in the inventory reference.

findings.json has this shape:

{
  "metadata": {
    "scan_id": "uuid",
    "provider": "AWS",
    "account_id": "...",
    "region": "...",
    "started_at": "...",
    "completed_at": "..."
  },
  "summary": {
    "total_assets": 0,
    "total_findings": 0,
    "total_edges": 0,
    "severity_breakdown": {"CRITICAL": 0, "HIGH": 0},
    "compliance_frameworks": ["CIS", "NIST-800-53"]
  },
  "assets": [],
  "findings": [],
  "compliance": [],
  "graph": {"nodes": [], "links": []}
}

assets, findings and compliance are lists of the models below, serialised with Pydantic's JSON mode. graph is D3 force-layout data. cloudg report -i accepts this same file back, so the format round-trips.

Data models

Everything flowing through cloudg is a Pydantic v2 model from cloudg.schema.models. All models allow extra fields, so scanner-specific attributes survive serialisation.

Finding

The unit every scanner and parser produces.

Field	Type	Notes
id	str	auto-generated UUID; reachability findings get a deterministic UUID5 of the rule and the asset's stable key (its ARN, else its asset ID), so the same finding keeps its id from scan to scan
resource_id	str	internal asset ID or native identifier
resource_arn	str | None	cloud-native identifier when known
severity	Severity	CRITICAL, HIGH, MEDIUM, LOW, INFO
title	str	display title; scanner tags such as [Checkov/terraform] are stripped during cross-scanner comparison
description	str
evidence	str | None	file path, package version, raw product fields
remediation	str | None
source_tool	str	"prowler", "checkov", ...; becomes "prowler, checkov" after a merge
source_finding_id	str | None	the scanner's own check or CVE ID; drives dedupe
compliance_frameworks	list[str]	unioned on merge
cvss_score	float | None	0.0 to 10.0
detected_at	datetime
is_suppressed	bool
risk_score	computed float	severity base averaged with CVSS when present

CloudAsset

Field	Type	Notes
id	str	auto UUID
arn	str | None	native identifier
name	str
asset_type	AssetType	156-value taxonomy covering compute, networking, storage, databases, IAM and identity federation, keys and secrets, logging, containers and Kubernetes, integration, data and ML platforms, DNS and deployment, security services and scanners, organization and governance, hybrid networking, OTHER; the full list with what maps to each is in the inventory catalog
provider	CloudProvider	AWS, AZURE, GCP
region	str	"global" for regionless resources
account_id	str | None
tags	dict[str, str]
metadata	dict[str, Any]	normalised extra attributes; inventory assets also carry declared relations and identifier aliases
collected_at	datetime	when the asset was built
is_internet_exposed	bool	set by the collector; the reachability analysis marks the graph nodes it reaches and reports them as findings, but does not change the asset
raw_data	dict[str, Any]	raw API payload, never serialised
display_id	computed str	arn or id

id is regenerated on every run; correlate assets across runs by arn.

NetworkEdge

Directed edge between two asset IDs (source_id, target_id); direction always reads "source verb target".

Field	Type	Notes
id	str	auto UUID
source_id, target_id	str	asset IDs; the far side of a network rule edge can be a CIDR (or an Azure service tag) instead of an asset ID
edge_type	EdgeType	coarse class, below
relationship	str | None	fine-grained ontology relation (TRIGGERED_BY, RUNS_ON, CROSS_ACCOUNT_TRUST, ENCRYPTED_BY_KMS, ...)
properties	dict[str, Any]	relation detail: trust conditions, granted actions, EKS access policies, notification events, cross_account, external_reference, hierarchy
description	str | None	human-readable explanation
ports, port_range, protocol, cidr, direction		network rule edges

edge_type values:

Group	Values
Network	SECURITY_GROUP_RULE, NACL_RULE, ROUTE, PEERING, LOAD_BALANCER_TARGET, INTERNET_EXPOSED
Structure	CONTAINS, ATTACHED_TO, REFERENCES
Identity	IAM_TRUST, IAM_POLICY_ATTACHMENT, ASSUMES_ROLE, GRANTS_ACCESS
Workload	INVOKES, USES_IMAGE, LOGS_TO
Security and governance	PROTECTS, MONITORS, MANAGES, GOVERNS

What each one means, its typical endpoints and how it counts for dependency analysis is in the inventory reference; every observed source type → relationship → target type combination is in the relationship matrix.

ComplianceResult

One control-level verdict: framework (for example CIS, NIST-800-53), control_id, control_title, status (PASS, FAIL, NOT_APPLICABLE, MANUAL) and the finding_ids behind it.

ScanResult

The aggregate container the normaliser returns and the renderers consume: assets, findings, edges, compliance, timing fields, and a computed summary dict with totals and the severity breakdown.

Configuration

config.yaml mirrors every CLI flag and adds persistent settings; pass it with -c or keep it next to where you run. Everything is optional, defaults are sensible. The shipped config.yaml documents each field inline. The structure, with defaults:

providers: [aws]                # aws, azure, gcp; several at once is fine

aws:
  regions: [us-east-1]          # or [all]
  accounts: []                  # multi-account fan-out
  role_name: null               # role to assume in each account
  role_arn: null                # single-account AssumeRole or OIDC target
  external_id: null
  web_identity_token_file: null
  profile: null
  max_retries: 10
  retry_mode: adaptive
  organization:                 # AWS Organizations / Control Tower (cloudg map)
    enabled: false
    role_name: null             # default AWSControlTowerExecution
    include_ous: []             # IDs, ARNs or names; nested OUs included
    exclude_accounts: []
    include_management_account: true
    include_suspended: false
    use_governed_regions: true  # regions [all] -> Control Tower governed regions
    control_tower: true
    home_region: null           # auto-detected
    map_structure: true         # org, OU, account, SCP, control nodes

azure:
  subscription_ids: []
  tenant_id: null
  client_id: null
  client_secret: null
  certificate_path: null
  federated_token_file: null
  use_managed_identity: false
  all_subscriptions: true       # every Enabled subscription when subscription_ids is empty
  map_management_groups: true   # management groups, subscriptions, Azure Policy
  regions: [ALL]

gcp:
  project_ids: []
  organization_id: null         # set to collect org-wide and map folders / org policy
  collection_scope: auto        # auto | organization | project
  skip_asset_types: [...]       # high-churn types (Pods, ReplicaSets, Events, ...) by default
  iam_policies: true
  map_hierarchy: true
  org_policies: true
  vpc_service_controls: true
  credentials_file: null
  impersonate_service_account: null
  regions: [ALL]

scanners:
  enabled: [prowler, scoutsuite, checkov, trivy, iam]
  timeout_seconds: 3600
  iac_directories: []           # targets for Checkov and Trivy fs
  trivy_images: []
  checkov_frameworks: []        # empty = auto-detect
  prowler_extra_args: []        # passthrough args, per scanner

inventory:
  tagging_sweep: true           # AWS catch-all sweep (Resource Groups Tagging API)
  link_references: true         # generic cross-service REFERENCES edges
  services: [all]               # families or collector names (see cloudg map)
  exclude_services: []
  kubernetes: true              # workloads inside EKS clusters
  kubernetes_timeout: 10
  iam_resource_edges: true      # principal -> granted resource edges
  max_images_per_repository: 20
  stack_resources: true         # CloudFormation stack -> resource edges
  account_hierarchy: true       # account nodes containing top-level resources
  cloud_control: true           # AWS breadth sweep over every listable resource type
  cloud_control_types: []       # only these CloudFormation types / prefixes
  cloud_control_exclude: []
  cloud_control_concurrency: 6

graph:
  persist_graphml: true
  compute_attack_paths: true

ontology:
  enabled: true
  export_formats: [turtle, json-ld]

rag:
  enabled: true
  chunk_strategy: hybrid        # entity | community | relation_group | hybrid
  max_chunk_tokens: 2000

terraform:
  enabled: false
  output_dir: ./reports/terraform

report:
  formats: [html, json, svg]
  inline_js: true               # air-gapped friendly

rulesets:
  rules_dir: null               # defaults to the packaged cloudg/rules/

ratelimit:                      # cloud API throttling resilience (0.6.0)
  enabled: true
  aws:                          # azure and gcp take the same keys
    max_rps: null               # null = built-in (AWS 20, Azure 15, GCP 10 per service)
    burst: null
    account_max_rps: null       # per account / subscription / project (Azure 20)
    account_burst: null         # (Azure 200)
    global_max_rps: null        # everything this process sends to the provider
    global_burst: null
    max_concurrency: null       # requests in flight per account (AWS 128, Azure 16, GCP 16)
    adaptive: true              # halve on throttling, recover additively
    min_rps: null
    breaker_threshold: 5
    breaker_cooldown_seconds: 60
    max_retries: 8              # Azure SDK retries, GCP retries per page (AWS: aws.max_retries)
    max_backoff_seconds: 60
    deadline_seconds: 900
    retry_budget: 500           # retries per provider, all SDKs, refilled by successes
    services: {}                # {ec2: {max_rps: 10, burst: 50}, "ec2.DescribeImages": {...}}
    live_cooldown_seconds: null
  live_cooldown_seconds: 120    # guard on live collections triggered through MCP
  live_max_concurrent: 1
  live_max_concurrent_total: 2
  live_caller_max_operations: 0
  live_caller_window_seconds: 3600
  live_operation_timeout_seconds: 3600  # cancel a live collection after this (null: no limit)

concurrency_limit: 5

aws.max_retries and aws.retry_mode configure botocore's own retries. They were accepted but ignored before 0.6.0; now they are applied to every AWS client. botocore counts max_retries after the first attempt, so the default of 10 allows up to 11 attempts per call. The ratelimit section is explained under Rate limits and throttling.

In code, the same structure is CloudGConfig, a Pydantic model:

from cloudg import CloudGConfig, load_config

config = load_config("config.yaml")          # file, validated
config = CloudGConfig(providers=["aws"])     # or programmatic
config.scanners.enabled = ["prowler", "iam"]
config.aws.regions = ["eu-west-1", "eu-central-1"]

Python API

CloudGEngine in cloudg.api is the integration entry point: async-first, sync wrappers included, built for embedding in larger systems (SIEM pipelines, orchestration platforms, command extensions such as hol-guard). The package root re-exports the essentials: CloudGEngine, CloudGConfig, load_config, PipelineResult, CollectionResult, AnalysisResult, InventoryMapper, InventoryResult.

The CLI is a thin layer; nearly everything cloudg does is public, importable API. Beyond the engine itself, the inventory mapping API and the deeper toolkit below are designed to be used directly from code.

Construction and hooks

from cloudg import CloudGConfig, CloudGEngine

engine = CloudGEngine(CloudGConfig(providers=["aws"]))

engine.on_phase_start = lambda phase: print(f"phase: {phase}")
engine.on_finding = lambda finding: siem.send(finding.model_dump(mode="json"))
engine.on_error = lambda phase, exc: alerting.notify(phase, exc)

Hook signatures:

Hook	Signature	Fires
on_phase_start	(phase: str) -> None	at each phase: collection, scanning, ingest, analysis, normalisation, reporting
on_collection_complete	(result: CollectionResult) -> None	after collection
on_finding	(finding: Finding) -> None	once per finding, after scan or ingest
on_scan_complete	(findings: list[Finding]) -> None	after scan or ingest
on_analysis_complete	(result: AnalysisResult) -> None	after analysis
on_error	(phase: str, exc: Exception) -> None	on any phase failure

Exceptions raised inside a hook are swallowed, so a broken callback cannot take the pipeline down.

Methods

run_pipeline(output_dir="./reports") -> PipelineResult and its sync twin run_pipeline_sync() run everything: collect, scan, analyse, normalise, report.

collect() -> CollectionResult runs multi-provider collection alone. The result carries assets (list[CloudAsset]), edges (list[NetworkEdge]), per-service coverage records, providers_scanned, regions_scanned and duration_ms.

scan(assets, edges, iac_dir=None, images=None, profile=None, output_dir="./reports") -> list[Finding] runs the scanners from config.scanners.enabled in parallel, plus graph reachability analysis and the IAM linter. Missing binaries are skipped.

ingest_reports(reports: dict[str, list[str | Path]]) -> list[Finding] parses existing scanner outputs instead of running anything. Keys are tool names (prowler, scoutsuite, checkov, trivy), values are lists of report paths, file or directory. Unparseable paths are logged and skipped rather than raising.

normalise_findings(findings, assets=None) -> ScanResult applies the full dedupe, cross-scanner merge, CVSS rescoring and compliance mapping. This is the same code path run_pipeline uses.

analyze(assets, edges, findings, output_dir="./reports") -> AnalysisResult builds the graph and produces the ontology, RAG chunks, Terraform files, attack paths and reachability findings, gated by the corresponding config sections.

run_from_reports(reports, output_dir="./reports") -> PipelineResult and run_from_reports_sync(...) chain ingest, normalise and report generation in one call.

map_inventory(output_dir=None, findings=None, tagging_sweep=None) -> InventoryResult and map_inventory_sync(...) run the scanner-independent inventory mapping: deep collection across all configured providers, catch-all sweeps, and relationship linking. Passing output_dir exports the map files; passing findings as well exports the merged asset and compliance maps.

Inventory mapping API

cloudg.inventory is a standalone package; the engine is optional:

from cloudg import CloudGConfig
from cloudg.inventory import InventoryMapper, RelationshipLinker

config = CloudGConfig(providers=["aws", "azure", "gcp"])
mapper = InventoryMapper(config)

inventory = mapper.map_inventory_sync()      # InventoryResult
print(inventory.summary)                     # services, types, regions, exposure
inventory.export("./reports")                # inventory-map.json / .graphml / graph

InventoryResult carries assets, edges, coverage, providers, regions, organization (the discovered topology, when mapped), unresolved_references and throttling (the run's throttling telemetry, or None when no API pushed back), plus a computed summary (totals, per-service/type/region/account breakdowns, edge types and relationships, cross-account edges, external accounts, security service gaps, internet-exposed and unlinked counts). export(dir) writes inventory-map.json, inventory-map.graphml, inventory-graph.json, inventory-dependencies.json and, for an organization, inventory-organization.json; InventoryResult.load(dir_or_file) reads a map back.

Interdependency questions go through DependencyGraph:

graph = inventory.dependency_graph()
role = graph.find("arn:aws:iam::123456789012:role/app-role")
needs = graph.depends_on(role.id)           # upstream, breadth-first
needed_by = graph.dependents(role.id)       # downstream: blast radius
view = graph.tree(role.id, direction="both", max_depth=3)
graph.shared_dependencies(top=25)
graph.blast_radius(top=25)
inventory.analysis()                        # everything above + coverage

Organizations and Control Tower discovery is usable on its own:

import boto3
from cloudg.inventory import discover_organization

topology = discover_organization(boto3.Session())
topology.governed_regions, topology.shared_accounts      # Control Tower
topology.target_accounts(include_ous=["Workloads"])     # nested OUs included
assets = topology.to_assets()                            # org, OUs, accounts, SCPs, controls

Merging scanner findings (from scan(), ingest_reports(), or a saved raw-findings.json) happens after the fact, so the mapper never touches the scanners:

asset_map = mapper.build_asset_map(inventory, findings)         # asset -> findings/severity
compliance_map = mapper.build_compliance_map(inventory, findings)  # framework -> assets
mapper.export_merged(inventory, findings, "./reports")          # asset-map.json + compliance-map.json

RelationshipLinker works on any list[CloudAsset], whoever collected them:

linker = RelationshipLinker(assets)
linker.seed_existing(existing_edges)   # don't duplicate edges you already have
edges = linker.link()                  # typed edges, see Relationships
linker.external_assets                 # placeholder nodes for unmapped accounts
linker.unresolved                      # declared references that resolved to nothing

Reference data lives in YAML catalogs under cloudg/inventory/catalogs/ rather than in code: aws_cloudcontrol.yaml (types the deep collectors already cover, account-wide types, CloudFormation type → asset type, per-type secret fields), aws_arn_types.yaml (ARN → asset type for sweep hits), azure_arm_types.yaml (ARM resource type → asset type) and gcp_asset_types.yaml (Cloud Asset Inventory type → asset type). Point $CLOUDG_CATALOG_DIR at a directory holding files of the same names to extend or override them without touching cloudg: mappings merge key by key, lists are extended. Unknown asset type names are rejected at load time.

Your own collectors can declare relationships the same way the built-in ones do, by listing them in metadata["relations"]: {"target": <any identifier>, "edge": "INVOKES", "relationship": "TRIGGERED_BY", "reverse": false}.

The deep collectors are public too, when you want single-region, single-account control: AWSDeepInventoryCollector(session, region="us-east-1", account_id=None, tagging_sweep=True, **options) (keyword options is_primary_region, services, exclude_services, kubernetes, kubernetes_timeout, iam_resource_edges, max_images_per_repository, stack_resources, cloud_control, cloud_control_types, cloud_control_exclude, cloud_control_concurrency; an unknown option raises TypeError), AzureDeepInventoryCollector(credential, subscription_id, graph_client_factory=None, use_resource_graph=True) and GCPDeepInventoryCollector(project_id, credentials=None, **options) (keyword options organization_id, scope, project_filter, skip_asset_types, page_size, include_iam, client, coverage, timeout, link_locally) all implement the standard collect() / collect_edges() / run() collector interface. Their assets carry declared relations; pass them through RelationshipLinker for the typed edges.

Every structure above, from summary to the dependency tree, is specified field by field in the inventory reference.

The deeper toolkit

Modules the pipeline uses internally that are equally useful standalone:

Module	Class	What it does from code
cloudg.graph.builder	GraphBuilder	build(assets, edges) → NetworkX DiGraph, with parallel security group, NACL and internet-exposure rules between the same two nodes merged into one edge (port_range and protocol become comma lists); find_attack_paths(src, dst), find_lateral_movement_paths(), compute_centrality() for blast-radius scoring; to_d3_json(), to_cytoscape_json(), save_graphml() / load_graphml()
cloudg.graph.reachability	ReachabilityAnalyzer, network_flow_graph, finding_id, asset_key	BFS from the internet over network-flow edges only. Entry points are 0.0.0.0/0, ::/0 and the Azure Internet, Any and * sources. The walk follows INTERNET_EXPOSED, ingress SECURITY_GROUP_RULE and NACL_RULE edges, ROUTE, PEERING and LOAD_BALANCER_TARGET source to target; ATTACHED_TO backwards into a security group, NSG or NACL and forwards from a network interface or Elastic IP; CONTAINS only from a VPC, VNet or subnet. It never follows IAM edges, and not INVOKES either, so a function behind a public API gateway is not flagged. Security groups, NSGs, NACLs and target groups are hops: they are marked exposed but get no exposure finding of their own. generate_findings() returns exposure and sensitive-port findings with deterministic ids; flow_hops(), flow_successors(), internet_entry_points() and network_flow_graph() expose the same walk; compute_blast_radius(node) follows every edge type
cloudg.graph.ports	parse_port_ranges, edge_port_ranges, port_in_ranges, is_internet_source	the port and edge rules the builder, the reachability analysis and the ontology share: port_range strings parsed into numeric ranges, filter rule and egress checks, internet sources
cloudg.graph.ontology	CloudOntology	build(assets, edges, findings) infers 64 typed RDF relations; save(path, fmt) writes Turtle/JSON-LD/XML; query the graph with SPARQL via rdflib
cloudg.graph.rag_export	RAGExporter	export_all(...) chunks the infrastructure three ways (entity, community, relation group) into JSONL for retrieval pipelines
cloudg.renderers.terraform_export	TerraformExporter	export(assets, edges) recreates live infrastructure as .tf.json plus an import.sh; preview(assets) reports mappable coverage first
cloudg.normaliser	FindingsNormaliser	the full dedupe / cross-scanner merge / CVSS rescore / compliance mapping pass, on any list[Finding]
cloudg.ingest	parse_report, ingest_reports	every scanner's parser, standalone; no engine, no credentials
cloudg.coverage	CollectionCoverage	per-service success/failure/asset-count records every collector produces
cloudg.inventory.dependencies	DependencyGraph, cross_account_edges, security_coverage	dependency walks, shared dependencies, blast radius, cross-account edges and security service coverage over any assets and edges
cloudg.inventory.linker	RelationshipLinker	resolve declared relations and identifiers across services, regions and accounts into typed edges
cloudg.inventory.organization	discover_organization, OrganizationTopology	AWS Organizations and Control Tower discovery, account selection, map nodes
cloudg.inventory.kubernetes	collect_eks_workloads, KubernetesReader	read-only Kubernetes object mapping for EKS clusters
cloudg.inventory.catalogs	load_catalog, asset_type_map	the YAML reference catalogs with $CLOUDG_CATALOG_DIR overlays
cloudg.registry	PluginRegistry	entry-point discovery of collector and scanner plugins

MCP layer API

cloudg.mcp exposes the inventory, graph, findings, compliance and ontology to AI agents over the Model Context Protocol; the MCP server chapter shows it in use. Importing it needs nothing beyond cloudg's own dependencies. Its exports:

Name	What it is
CloudGMCPLayer	The layer: registry, workspace, policy and middleware. call_tool, read_resource, get_prompt, complete, list_* and *_wire methods, register_into(server), serve(transport)
default_registry(), Registry	The built-in catalog, and the container for your own tools (@registry.tool, @registry.resource, @registry.resource_template, @registry.prompt)
Workspace, Dataset	Loaded datasets, the active one, the allowed directories
Policy, available_profiles()	Access rules, rate limits and transform pipelines; the seven built-in profile names
Principal, ToolContext, CallInfo	Who is calling, what a tool receives, what middleware sees
register_into(layer, server), build_server(layer, flavor), export_definitions(layer)	Mount into an MCP SDK or fastmcp server, build a standalone server object, or get framework-free wire definitions plus async handlers
Pipeline, TokenVault, TransformContext, build_transform()	The transform machinery and the pseudonym vault
ToolSpec, ResourceSpec, ResourceTemplateSpec, PromptSpec, PromptArgument, ToolAnnotations, Capability, Sensitivity, Icon	Primitive specifications
ToolResult, TextContent, ImageContent, EmbeddedResource, ResourceLink, TextResourceContents, ContentAnnotations, PromptMessage, PromptResult	Results and content blocks
MCPLayerError, AccessDeniedError, InvalidArgumentsError, NotFoundError, RateLimitedError	Errors, each with its JSON-RPC code

CloudGMCPLayer also takes all of its keyword arguments as one LayerOptions object (from cloudg.mcp.layer import LayerOptions; CloudGMCPLayer(options=LayerOptions(policy="strict", prefix="cloudg_"))), and an unknown keyword raises TypeError. A tool handler declares its arguments as parameters after ctx, or, when there are many, as one pydantic model in a parameter named args (def my_tool(ctx, args: MyArgs)); both forms produce the same flat input schema.

The server side lives in cloudg.mcp.server (serve, serve_async, ServeOptions, create_layer_from_options; serve_async(layer, transport, options=ServeOptions(...)) takes the serve settings as one object), the CLI in cloudg.mcp.cli (mcp_group, main, client_config), the middleware in cloudg.mcp.middleware (AuditLogMiddleware, MetricsMiddleware, CachingMiddleware, ConcurrencyLimitMiddleware, RetryMiddleware). Every signature is in the MCP guide.

Resilience API

cloudg.resilience is the throttling layer every collector goes through, and you can put your own cloud calls through it too (Rate limits and throttling):

Name	What it does
call_with_resilience(fn, *args, scope=..., policy=None, **kwargs)	Async. Breaker, deadline, shared rate limit, bulkhead, then the call; retries throttling and transient errors with decorrelated jitter, never sooner than Retry-After. Sync functions run in a worker thread
call_with_resilience_sync(...)	The blocking variant, for worker threads
@resilient(scope, policy=None)	Decorator form; scope may be a callable that receives the function's arguments
Scope(provider, account, region, service, operation), RetryPolicy(...)	Where a call goes; per-call retry bounds
classify(exc), ErrorKind, retry_after(exc), describe_error(exc), is_throttle(exc)	Provider-aware error classification (THROTTLED, TRANSIENT, FATAL) and the server's requested delay
classify_strict(exc), is_provider_answer(exc)	Classification from error codes, exception classes and HTTP statuses only, never message text (None when they say nothing); whether the error is an answer from the provider rather than a local failure
get_governor(), configure(config), reset_governor()	The process-wide limiter, breakers, budgets and stats
stats_scope(), current_stats(), ResilienceStats	Per-run throttling telemetry; get_governor().summary(stats) is what InventoryResult.throttling holds
LiveOperationGuard, GuardSettings, operation_key(), LiveOperationRejected (CooldownActive, LiveOperationBusy, CallerQuotaExceeded), LiveOperationTimeout	Single-flight, concurrency caps, cooldowns, caller quotas and an operation timeout for on-demand collections
RateLimiter, TokenBucket, LimitSpec, ProviderLimits, Bulkhead, CircuitBreaker, BreakerRegistry, DEFAULT_LIMITS	The building blocks
CircuitOpenError, DeadlineExceededError, RetryBudgetExhaustedError	cloudg's own errors, carrying retry_after

import asyncio

import boto3
from cloudg.resilience import Scope, call_with_resilience, stats_scope


async def main():
    ec2 = boto3.client("ec2", region_name="eu-west-1")
    scope = Scope("aws", "123456789012", "eu-west-1", "ec2", "DescribeVpcs")
    with stats_scope() as stats:
        vpcs = await call_with_resilience(ec2.describe_vpcs, scope=scope)
    print(len(vpcs["Vpcs"]), stats.messages())


asyncio.run(main())

cloudg.retry.with_retry keeps its signature and retries what it always did. It now also retries provider throttling that classify_strict recognises from an error code, exception class or HTTP status (Azure 429s, GCP ResourceExhausted), and never sleeps less than the server's Retry-After, capped at its max_delay. It does not retry cloudg's own resilience errors, a plain HTTP 500, or an error whose message merely mentions rate limiting.

PipelineResult

Field	Type
assets, edges, findings	the collected and normalised data
scan_result	ScanResult | None
graph_nodes, graph_edges, ontology_triples	int
rag_chunks_path, terraform_paths, report_paths	output locations
attack_paths	lateral movement paths from the graph
providers_scanned, regions_scanned, coverage	collection metadata
duration_ms, errors	run metadata
total_assets, total_findings, severity_breakdown	computed properties
to_summary()	one dict with the headline numbers, ready to serialise

Integration recipes

Working examples against the real API. Each is self-contained.

Aggregate existing scan results, get reports

from cloudg import CloudGConfig, CloudGEngine

engine = CloudGEngine(CloudGConfig())
result = engine.run_from_reports_sync(
    {
        "prowler": ["./prowler-output/"],
        "trivy": ["./trivy-image.json", "./trivy-fs.json"],
    },
    output_dir="./reports",
)

print(result.total_findings, result.severity_breakdown)
print(result.report_paths["html"])   # ./reports/report.html
for f in result.findings:
    if f.severity.value in ("CRITICAL", "HIGH"):
        print(f.risk_score, f.title, f.resource_arn)

Parse first, decide later

from cloudg import CloudGConfig, CloudGEngine
from cloudg.ingest import parse_report, ingest_reports

findings = parse_report("prowler", "./prowler-output/")
findings += parse_report("checkov", "./results_json.json")
# or in one call:
findings = ingest_reports({"prowler": ["./out/"], "checkov": ["results_json.json"]})

engine = CloudGEngine(CloudGConfig())
scan_result = engine.normalise_findings(findings)

print(scan_result.summary)
for c in scan_result.compliance:
    if c.status.value == "FAIL":
        print(c.framework, c.control_id, len(c.finding_ids))

The per-scanner parsers are also public, if you want a single tool with no engine:

from cloudg.scanners.trivy import TrivyScanner

findings = TrivyScanner.parse_report("./trivy-image.json")

Stream findings into a SIEM as they arrive

import asyncio
from cloudg import CloudGConfig, CloudGEngine

engine = CloudGEngine(CloudGConfig(providers=["aws"]))
engine.on_finding = lambda f: queue.put(f.model_dump(mode="json"))
engine.on_error = lambda phase, exc: log.warning("cloudg %s failed: %s", phase, exc)

result = asyncio.run(engine.run_pipeline(output_dir="/var/lib/cloudg/reports"))

Run the phases yourself

import asyncio
from cloudg import CloudGConfig, CloudGEngine

async def main():
    engine = CloudGEngine(CloudGConfig(providers=["aws"]))

    collection = await engine.collect()
    findings = await engine.scan(collection.assets, collection.edges)
    analysis = await engine.analyze(collection.assets, collection.edges, findings)
    scan_result = engine.normalise_findings(findings, assets=collection.assets)

    print(analysis.graph_nodes, "nodes,", len(analysis.attack_paths), "attack paths")
    return scan_result

asyncio.run(main())

Combine live collection with ingested reports

Collection needs credentials; ingest does not. Together they give you the graph outputs and the aggregated findings in one result:

import asyncio
from cloudg import CloudGConfig, CloudGEngine

async def main():
    engine = CloudGEngine(CloudGConfig(providers=["aws"]))

    collection = await engine.collect()
    ingested = engine.ingest_reports({"prowler": ["./ci-prowler-output/"]})
    scan_result = engine.normalise_findings(ingested, assets=collection.assets)
    await engine.analyze(collection.assets, collection.edges, scan_result.findings)

asyncio.run(main())

Map the inventory, overlay findings later

Inventory mapping needs read credentials and nothing else. Scanners can run elsewhere (CI, a schedule, another host) and merge in whenever their output arrives:

from cloudg import CloudGConfig, CloudGEngine
from cloudg.inventory import InventoryMapper

config = CloudGConfig(providers=["aws"])
config.aws.regions = ["ALL"]

engine = CloudGEngine(config)
inventory = engine.map_inventory_sync(output_dir="./reports")   # no scanners

print(inventory.summary["assets_by_service"])   # what exists
print(inventory.summary["internet_exposed"])    # what faces the internet
print(inventory.summary["unlinked_assets"])     # what nothing points at

# hours or days later, when scanner output exists:
findings = engine.ingest_reports({"prowler": ["./ci-prowler-output/"]})
mapper = InventoryMapper(config)
mapper.export_merged(inventory, findings, "./reports")
# -> asset-map.json (asset -> risk), compliance-map.json (framework -> assets)

Link relationships into someone else's inventory

The linker is pure post-processing. Feed it assets from any source that produces CloudAsset objects (a plugin collector, a CMDB import, a previous run):

from cloudg.inventory import RelationshipLinker
from cloudg.graph.builder import GraphBuilder

edges = RelationshipLinker(assets).link()
graph = GraphBuilder().build(assets, edges)
print(graph.number_of_nodes(), graph.number_of_edges())

Gate a CI job on severity

import sys
from cloudg import CloudGConfig, CloudGEngine

result = CloudGEngine(CloudGConfig()).run_from_reports_sync(
    {"checkov": ["results_json.json"], "trivy": ["trivy.json"]}
)
critical = result.severity_breakdown.get("CRITICAL", 0)
if critical:
    print(f"{critical} critical findings, failing the build")
    sys.exit(1)

MCP server

cloudg 0.6.0 ships a Model Context Protocol layer, cloudg.mcp. An AI agent (Claude Desktop, Claude Code, Cursor, VS Code, or your own) can load an inventory map or a findings report, search and inspect assets, walk the relationship graph, ask for attack paths and blast radius, review findings and compliance gaps, query the ontology, and, when the policy allows it, start live collections. Every request goes through one policy that decides who may call what and transforms the data on its way out: secrets redacted, attacker-controllable text fenced, identifiers pseudonymised under the stricter profiles.

The full guide is MCP.md; this chapter is the summary.

Installing

pip install cloudg            # the layer, the CLI and the dependency-free native server
pip install "cloudg[mcp]"     # adds the official mcp SDK: the sdk flavor and register_into(...)
pip install fastmcp           # only to serve with fastmcp or mount into a fastmcp server

The mcp extra pins mcp>=1.30,<3; this release was verified with mcp 1.30.0 and 2.3.0, and with fastmcp 2.14.7, 3.4.8 and 4.0.11. Mounting into an SDK MCPServer (mcp 2.x) or FastMCP (mcp 1.x) goes through private SDK attributes that a future release can rename, so applications that use that path should pin the minor version they tested (for example mcp>=2.3,<2.4). Mounting into the SDK's low-level Server and the native server do not depend on those attributes. cloudg[all] includes the mcp extra.

Serving it

# stdio, for a desktop client that launches the server itself
cloudg mcp serve --dataset prod=./reports/inventory-map.json

# print the client configuration to paste (claude-desktop, claude-code, cursor, vscode)
cloudg mcp config --client claude-desktop --dataset prod=./reports/inventory-map.json

# streamable HTTP on 127.0.0.1:8765/mcp, with a bearer token mapped to a role
cloudg mcp serve --transport http --auth-token env:CLOUDG_MCP_TOKEN:analyst

The same commands are available as cloudg-mcp ... and python -m cloudg.mcp .... The server speaks both MCP protocol eras: the handshake versions 2024-11-05, 2025-03-26, 2025-06-18 and 2025-11-25, and the stateless 2026-07-28. --flavor picks the implementation: native (standard library only), sdk (the official SDK, needs cloudg[mcp]), fastmcp, or auto, which uses the SDK when it is installed and the native server otherwise. HTTP binds to loopback and checks Origin and Host by default; when binding anything else, pass --auth-token, --allowed-host and --allowed-origin.

Embedding it

Mount every tool, resource and prompt into an MCP server you already run. Here an mcp 2.x MCPServer with one tool of its own; an mcp 1.x FastMCP, a low-level Server and a fastmcp.FastMCP work the same way:

import asyncio

from mcp import Client
from mcp.server.mcpserver import MCPServer

from cloudg.mcp import CloudGMCPLayer

server = MCPServer("platform-tools")


@server.tool()
def ticket_count(team: str) -> int:
    """The host's own tool."""
    return 7


layer = CloudGMCPLayer(policy="standard")
layer.workspace.load("inventory/inventory-map.json", "prod")
layer.register_into(server, prefix="cloudg_")


async def main() -> None:
    async with Client(server) as client:
        tools = (await client.list_tools()).tools
        print(len(tools), [t.name for t in tools][:3])
        res = await client.call_tool("cloudg_count_assets", {"group_by": "provider"})
        print(res.structured_content["groups"])


asyncio.run(main())

74 ['ticket_count', 'cloudg_workspace_status', 'cloudg_list_datasets']
{'AWS': 36, 'AZURE': 5, 'GCP': 3}

Mount after the host has registered its own handlers: the host's tools stay, and names cloudg does not own still reach the host. Or skip MCP on the wire and call the layer in-process:

import asyncio
from pathlib import Path

from cloudg.mcp import CloudGMCPLayer, Workspace


async def main() -> None:
    layer = CloudGMCPLayer(policy="strict", workspace=Workspace(allowed_roots=[Path(".")]))
    await layer.call_tool("load_dataset", {"path": "inventory/inventory-map.json", "name": "prod"})
    res = await layer.call_tool("find_assets", {"internet_exposed": True, "limit": 2,
                                                "fields": ["name", "type", "arn"]})
    print(res.is_error, res.structured["total"], res.structured["items"])


asyncio.run(main())

False 6 [{'id': 'res-04ddd3e8be', 'name': 'res-04ddd3e8be', 'type': 'EC2', 'arn': 'arn:aws:ec2:us-east-1:934619913841:instance/res-c6e5dce4cb'}, ...]

Under strict the names, account ids and ARNs come back pseudonymised; passing a pseudonym back as an argument works, because the input pipeline reverses it. Set CLOUDG_MCP_VAULT_KEY to keep pseudonyms stable across restarts (without it they change with every process, as in this run). A policy with vault.path also saves the vault when the server shuts down, so issued pseudonyms survive a restart. Both examples ran against the synthetic estate in tests/mcp/fixtures/sample_estate.py.

CLI

Command	What it does
cloudg mcp serve	Run the server: --transport stdio/http/sse, --host, --port, --path, --flavor, --auth-token, --allowed-origin, --allowed-host, --cache-ttl, --max-concurrency, --log-level
cloudg mcp tools / resources / prompts	List what a principal sees under a policy (--role, --principal-id, --json for the exact wire definitions)
cloudg mcp call TOOL --args JSON	Call a tool in-process through the full policy; exit status 1 on a tool error
cloudg mcp read URI	Read a resource in-process
cloudg mcp config	Print a client configuration snippet

All of them except config accept --policy, --dataset [NAME=]PATH (repeatable), --prefix, --include-category / --exclude-category, --include-tool / --exclude-tool, --read-only (drops tools that touch the cloud, run processes or write files), --audit-log FILE, --timeout and --registry MODULE:ATTR.

Catalog

The built-in catalog has 74 tools, 13 resources, 11 resource templates and 11 prompts. The CLI adds cloudg://metrics. Tools by category:

Category	Tools	What they do
graph	17	neighbours, paths, exposure, attack and lateral paths, dependencies, blast radius, centrality, sub-graphs
inventory	11	search, inspect and aggregate assets, accounts, regions, tags, coverage, organization
findings	10	browse, summarise, prioritise, suppress and ingest security findings
workspace	8	load, select, snapshot, diff and unload datasets
ontology	6	RDF ontology, read-only SPARQL, semantic neighbourhoods, RAG chunks
compliance	5	framework posture, controls and gaps
live	5	the four live tools (map_inventory, collect_assets, run_scanners, run_pipeline) and rate_limit_status
privacy	5	policy, pseudonym vault and data-handling controls
meta	4	server capabilities and cloudg's vocabulary
export	3	Terraform recreation, reports, ontology files

The 11 prompts are packaged workflows such as security_posture_review, attack_surface_report, blast_radius_assessment, compliance_gap_analysis and incident_triage. Every primitive is documented in MCP_TOOLS.md.

The cloudg://docs resources let an agent read this reference and the inventory reference. The wheel ships both files inside the package (cloudg/_docs/), so they work from a plain pip install; a source checkout reads them from docs/, and CLOUDG_DOCS_DIR points elsewhere. The sdist includes every docs/*.md.

Policies and privacy

--policy takes a built-in profile, a YAML or JSON file, or inline JSON; the default is standard.

Profile	In short	Tools listed for the local user
open	no restrictions, no transforms; for trusted local use	74
standard	redacts secrets, drops private keys, masks credential ids, fences attacker-controllable text, refuses secrets in arguments; reveal_token only for admin and privacy-admin	73
strict	standard plus pseudonymised identifiers, IPs, e-mails and tag values; hides the live, export and RESTRICTED tools (among them sparql_query, subgraph_export and the text graph and ontology exports); rate limits on every tool	61
read_only	standard without cloud access, processes or file writes	66
airgapped	standard without cloud access or processes	69
audit	read_only with every call logged and risk annotations inline	66
soc-analyst	a role-based example: analysts get pseudonymised data, a lead role sees it in clear	62; analyst 60, lead 70

Policies can extend each other, carry per-role rules, sensitivity ceilings and per-principal rate limits, and compose transforms (redact, sanitize, project, annotate, substitute, alias, regex_replace, rename_keys, template, guard_secrets, depseudonymize). The pseudonym vault (TokenVault) derives format-preserving pseudonyms with HMAC-SHA256. Every profile, key and transform is in MCP_PRIVACY.md. reveal_token (for the roles allowed to reveal) turns a pseudonym, or an alias a substitute or alias transform introduced, back into the real value.

Live tools

map_inventory, collect_assets, run_scanners and run_pipeline call cloud APIs with the server's own credentials, whoever the caller is, so restrict them to trusted roles. The strict, read_only, airgapped, audit and soc-analyst profiles hide them, and --read-only drops them. Before any cloud call, a credential preflight (preflight=true by default, at most 25 seconds) checks that the configured AWS, Azure and GCP credentials work, so a missing profile fails in seconds instead of after every collector has retried.

Then the call goes through the live operation guard. Identical concurrent calls share one run; a scope collected in the last two minutes is refused with retry_after_seconds and a pointer to a loaded live dataset that covers the same providers; at most one live operation runs per scope and two overall. A refusal is a tool error with code -31029 and reason cooldown, busy or quota. Principals with the admin or operator role can pass force=true to skip the cooldown. An operation still running after ratelimit.live_operation_timeout_seconds (default 3600) is cancelled and frees its scope. The answer of every live tool includes a throttling block when a cloud API pushed back during the run. rate_limit_status and the resource cloudg://ratelimit show the cooldowns, the operations in flight and the throttling state without touching any cloud API. The guard's settings are the ratelimit.live_* keys described in the next chapter.

Rate limits and throttling

Since 0.6.0 every cloud API call cloudg makes goes through cloudg.resilience, one process-wide governor shared by all clients:

• Token buckets per account x region x service (per API action for EC2; per subscription and resource provider on Azure; per project or organization and RPC on GCP), with defaults at or below each provider's published limits: AWS 20 rps per service and 20 per EC2 action, Azure 20 rps per subscription and 3 per user for Resource Graph, GCP 1.5 rps for Cloud Asset ListAssets.
• Adaptive rates: a throttled scope's rate is halved for every client using it, paused for the server's Retry-After, and recovered by 5% of the ceiling every 5 seconds without throttling.
• Concurrency caps per account (max_concurrency: 128 requests in flight on AWS, 16 on Azure and GCP).
• Circuit breakers per service, or per API action for EC2: after 5 consecutive calls that still failed after retries, calls are skipped for 60 seconds, then one probe goes through. A throttled DescribeSubnets does not block DescribeVpcs.
• Retries with decorrelated jitter that never undercut Retry-After, bounded by attempts, a per-call deadline and a per-provider retry budget that every SDK draws from. botocore keeps doing the retrying on AWS (aws.retry_mode, aws.max_retries), azure-core on Azure (an explicit RetryPolicy with cloudg's retries, backoff cap and timeout), google-api-core on GCP (a Retry capped per page); all of them report back to the governor. The first two retries of every call are free; after that each retry draws on the budget, and when it is spent the call gives up.

Throttling never fails a map. A service that stayed throttled is recorded in coverage as FAILED or PARTIAL with an error starting throttled:, the rest of the map continues, and the run's numbers are attached to the result:

result = InventoryMapper(config).map_inventory_sync()
result.throttling                    # None, or totals / messages / skipped / scopes / breakers / slowed_buckets
result.summary["throttling"]         # totals, messages, skipped

inventory-map.json carries the full block under throttling, and cloudg map logs each message, for example Throttling: aws/123456789012/us-east-1/ec2: throttled 11x, slowed to 5.0 rps, 10 retries, 1 calls gave up.

Everything is configured under ratelimit (structure under Configuration). Unset limits fall back to the built-ins, and service overrides merge into them:

ratelimit:
  aws:
    account_max_rps: 40                       # all services of one account together
    services:
      resourcegroupstaggingapi: {max_rps: 1, burst: 2}
      ec2.DescribeImageAttribute: {max_rps: 5, burst: 10}
  azure:
    services:
      resourcegraph: {max_rps: 1, burst: 5}
  live_cooldown_seconds: 600                  # MCP live collections, per scope

A misspelled key anywhere under ratelimit is logged as Unknown config key ratelimit.aws.max_rpss is ignored and the rest still loads. ratelimit.enabled: false turns it all off (botocore's own retries remain), and ratelimit.<provider>.enabled: false does it for one provider. The rate limits and throttling guide has the architecture, the per-provider wiring, the built-in limits with their sources, every key with its default and where it takes effect, example output of a throttled run, the Python API and tuning advice.

Plugin system

Custom collectors and scanners register through setuptools entry points; no core changes needed. PluginRegistry (in cloudg.registry) discovers them at startup and falls back to the built-ins for any name not overridden.

A minimal scanner plugin:

# my_package/scanner.py
from cloudg.schema.models import Finding, Severity

class MyScanner:
    def __init__(self, **kwargs):
        ...

    def run(self) -> list[Finding]:
        return [
            Finding(
                resource_id="arn:aws:s3:::example",
                severity=Severity.HIGH,
                title="Example finding",
                description="What was found and why it matters",
                source_tool="myscanner",
                source_finding_id="MY_CHECK_001",
            )
        ]

Registered in the plugin's own pyproject.toml:

[project.entry-points."cloudg.scanners"]
myscanner = "my_package.scanner:MyScanner"

[project.entry-points."cloudg.collectors"]
mycloud = "my_package.collector:MyCollector"

After pip install, the name is available in scanners.enabled and --scanners. Collectors follow the base interface in cloudg/collectors/base.py. A collector plugin's assets join the inventory map like the built-in ones: declare what each asset talks to in metadata["relations"] and extra identifiers in metadata["aliases"], and the relationship linker resolves them (see Inventory internals). If your scanner emits a stable check ID in source_finding_id, you can add it to cloudg/rules/check_equivalence.yaml so its findings merge with equivalent checks from other tools.

Authentication

Every provider supports several methods, resolved in a fixed priority order, so the same config works on a laptop, in CI and on cloud compute.

AWS, in priority order: direct keys (--aws-key, --aws-secret, --aws-session-token or the standard env vars); OIDC web identity federation (--aws-role-arn with --aws-web-identity-token-file, the GitHub Actions, GitLab CI and EKS pattern with no long-lived keys); a named CLI profile via --profile, SSO included; or nothing, letting the default chain pick up env vars, cached SSO or the instance role. On top of any of these, STS role assumption with --aws-role-arn and, for the third-party auditor pattern, --aws-external-id. Multi-account fan-out uses accounts plus role_name in the config, assuming that role in each account before collecting; the caller's own account is collected with the base credentials.

AWS Organizations and Control Tower: run cloudg map --org with credentials for the management account (Organizations APIs also work from a delegated administrator; the Control Tower APIs need the management account). cloudg lists the org read-only (organizations:Describe*/List*, controltower:List*/Get*) and then assumes the member role in each selected account. AWSControlTowerExecution exists in every account Control Tower enrolled, but it carries administrator permissions; the aws-controltower-ReadOnlyExecutionRole cannot be used directly because it only trusts a Lambda-only role in the audit account. For production, deploy a dedicated read-only role (for example cloudg-readonly with the SecurityAudit and ViewOnlyAccess managed policies, trusting the management account) to all accounts with a service-managed CloudFormation StackSet targeting your OUs, and pass it with --org-role cloudg-readonly. Mapping Kubernetes workloads additionally needs an EKS access entry for that role with AmazonEKSViewPolicy.

Azure, in priority order: workload identity federation (--azure-tenant-id, --azure-client-id, --azure-federated-token-file); a service principal with a client secret (--azure-client-secret) or a certificate (--azure-cert-path); managed identity (--azure-managed-identity, with managed_identity_client_id in config for user-assigned identities); or the DefaultAzureCredential chain, which covers az login.

GCP: a credentials file via --gcp-credentials-file (service account key JSON or a workload identity federation external_account config), or application default credentials. --gcp-impersonate-sa layers service account impersonation on either.

Docker

The image bundles all four scanner binaries, so it is the shortest path to the full pipeline:

docker pull ghcr.io/morpheuslord/cloudg:latest
docker compose run --rm cloudg run -p aws --regions us-east-1

Or build locally with docker build -t cloudg:latest .

---

Found a gap in this reference? Open an issue.